On Bayesian Predictive Efficiency Rankings (DSMok1, 2011)
Posted: Fri Apr 15, 2011 1:04 am
recovered pages 1 & 2 of 3
Author Message DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 5:17 pm Post subject: On Bayesian Predictive Efficiency Rankings
I just put up a tremendously long post on my website discussing how to use Bayesian updating with NBA predictive efficiency ratings. I'm sure there are some flaws (such as how to adjust for opponent), so let me know! http://godismyjudgeok.com/DStats/2011/n ... com/DStats Twitter.com/DSMok1
Back to top
back2newbelf
Joined: 21 Jun 2005
Posts: 274
Posted: Tue Feb 15, 2011 5:27 pm Post subject:
I get an Error in query: User not signed in <a target="_blank" href="https://spreadsheets0.google.com/">Sign in</a> and the last charts don't show up_________________http://stats-for-the-nba.appspot.com/
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 5:36 pm Post subject:
back2newbelf wrote:
I get an Error in query: User not signed in <a target="_blank" href="https://spreadsheets0.google.com/">Sign in</a> and the last charts don't show up
Try again--does it work now?_________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
back2newbelf
Joined: 21 Jun 2005
Posts: 274
Posted: Tue Feb 15, 2011 5:50 pm Post subject:
Yes it works now_________________http://stats-for-the-nba.appspot.com/
Back to top
Ilardi
Joined: 15 May 2008
Posts: 265
Location: Lawrence, KS
Posted: Tue Feb 15, 2011 6:38 pm Post subject:
DSM - Intriguing work, as usual. So, I'm curious to know if your Bayesian model does a reasonable job approximating the observed distribution of actual game results? In other words, if we took a largish historical sample of n games, and for each game computed your model's predicted game outcome (based on the Bayesian efficiency differential of each team's antecedent games), would the actual observed game outcomes be distributed more or less normally, with the predicted distribution means and sd's? Steve
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 6:57 pm Post subject:
Ilardi wrote:
DSM - Intriguing work, as usual. So, I'm curious to know if your Bayesian model does a reasonable job approximating the observed distribution of actual game results? In other words, if we took a largish historical sample of n games, and for each game computed your model's predicted game outcome (based on the Bayesian efficiency differential of each team's antecedent games), would the actual observed game outcomes be distributed more or less normally, with the predicted distribution means and sd's? Steve
Sounds like another post! _________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Crow
Joined: 20 Jan 2009
Posts: 821
Posted: Tue Feb 15, 2011 9:07 pm Post subject:
Do you want to predict the playoffs with this tool and compare to Hollinger's and the others out there?
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 11:33 pm Post subject:
Crow wrote:
Do you want to predict the playoffs with this tool and compare to Hollinger's and the others out there?
Not really. I'd prefer to predict the playoffs based on Bayesian-based ASPM._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 70
Posted: Wed Feb 16, 2011 12:44 am Post subject:
Not sure if you would consider this helpful or not DSM, but I know a programmer who used Bayesian priors for all boxscore stats in a predictive format two seasons ago (and had a methodology for ranking the importance of each stat based on its relationship to winning as the season played itself out). If I recall, he was surprised to find that the most predictive hunk of games was the prior five...rather than much longer term samples. He had expected larger sample sizes to make better predictions. Ultimately he decided the approach was still inferior to the prediction markets reflected by the Vegas pointspreads. Those have the ability to react to on the fly information like injuries, magnified fatigue spots, guys coming back from injuries, etc... Best of luck with your efforts...
Back to top
bbstats
Joined: 25 Apr 2010
Posts: 46
Posted: Wed Feb 16, 2011 12:48 am Post subject:
You had me at "bayesian." Great stuff._________________http://thebasketballdistribution.blogspot.com http://twitter.com/bbstats
Back to top
bbstats
Joined: 25 Apr 2010
Posts: 46
Posted: Wed Feb 16, 2011 1:08 am Post subject:
PS - Would I be correct in saying that the theory here is similar to Dean Oliver's Kalman Filter?_________________http://thebasketballdistribution.blogspot.com http://twitter.com/bbstats
Back to top
EvanZ
Joined: 22 Nov 2010
Posts: 295
Posted: Wed Feb 16, 2011 6:34 am Post subject:
bbstats wrote:
You had me at "bayesian." Great stuff.
He had me at "On...". Any time you see that leading preposition, you know some mathematical wizardry is to follow. BTW, has anyone used bootstrapping to do something similar? Or maybe that could be useful to create a prior?_________________http://www.thecity2.com http://www.ibb.gatech.edu/evan-zamir
Back to top
John Hollinger
Joined: 14 Feb 2005
Posts: 175
Posted: Wed Feb 16, 2011 10:48 am Post subject:
Fascinating stuff, and interesting that it supports my notion that Philly and Memphis are both a hell of a lot better than people realize and Utah is much worse. I suspect our big difference on Dallas is due to the timing of Dirks injury since I just weight last 10 rather than using a gradual function like you did....
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Wed Feb 16, 2011 11:23 am Post subject:
John Hollinger wrote:
Fascinating stuff, and interesting that it supports my notion that Philly and Memphis are both a hell of a lot better than people realize and Utah is much worse. I suspect our big difference on Dallas is due to the timing of Dirks injury since I just weight last 10 rather than using a gradual function like you did....
When you do the time-weighting, how & when in the calculations do you adjust for opponents? In this analysis, I first ran fully adjusted team efficiency differentials (over the whole season, w/o time-weighting), and used that value to pre-adjust the game efficiency differentials, before doing the Bayesian time-weighting. How do you adjust for opponents in the "last 10" component? I'm puzzling how to do that part of the analysis more rigorously._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 70
Posted: Wed Feb 16, 2011 1:12 pm Post subject:
Regarding JH's notes on Philly/Memphis/Utah, confirmation from the markets... *Philadelphia has topped expectations to the tune of 34-20-1 this season, including 29-14 the last 43 games. *Memphis is 35-21-1, including 28-12 the last 40 games. *Utah is 25-30-1 for the year, including 8-19 the last 27 games. If one accepts the premise that market prices are a composite of "what people are thinking," we can probably move that from notion to confirmed fact I'd think.
Author Message Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Thu Feb 17, 2011 12:26 am Post subject:
Updating to 35-20-1 for Philly vs. expectations, 30-14 the last 44 heading into the ASB after Wednesday's win in Houston... And, 25-31-1 for Utah, 8-20 the last 28 heading into the ASB after Wednesday's home loss to Golden State.
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Fri Feb 18, 2011 5:47 pm Post subject:
Now, with NCAA ratings! I also put up a spreadsheet which you all may find very interesting to manipulate._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 07, 2011 7:44 pm Post subject:
DSMok1 wrote:
Now, with NCAA ratings! I also put up a spreadsheet which you all may find very interesting to manipulate.
I've updated this data and added some more things. I still haven't gotten around to doing the standard errors for the NCAA and verifying the errors with game results, sorry. http://godismyjudgeok.com/DStats/2011/n ... com/DStats Twitter.com/DSMok1
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Sat Mar 12, 2011 10:10 pm Post subject:
And another new version, this time incorporating previous year information as a better-informed Bayesian prior: http://godismyjudgeok.com/DStats/2011/n ... com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Sun Mar 13, 2011 11:43 am Post subject:
DSM, would you consider the numbers in your Bayesian column to be a reasonable estimation of the point differences between teams in a 40 minute game? Meaning, Kansas is about 3 points better on a neutral court than Texas because they're 3 points higher? If not, is there a way to easily convert your output into something that would resemble a point differential scale so readers could compare it to the market prices that go up in the NCAA Tournament? Living in Austin, it was cool to see Texas so high. Wish they could find a high level of form more consistently. I'm afraid they peaked too early again...
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 6:29 am Post subject:
Jeff Fogle wrote:
DSM, would you consider the numbers in your Bayesian column to be a reasonable estimation of the point differences between teams in a 40 minute game? Meaning, Kansas is about 3 points better on a neutral court than Texas because they're 3 points higher? If not, is there a way to easily convert your output into something that would resemble a point differential scale so readers could compare it to the market prices that go up in the NCAA Tournament? Living in Austin, it was cool to see Texas so high. Wish they could find a high level of form more consistently. I'm afraid they peaked too early again...
Those are points per 100 possessions, not per game. To calculate the number of possessions expected in a game, take adjusted pace from Pomeroy for each team, and calculate as PaceA*PaceB/NCAAAvg. to get the expected pace for each game._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Ilardi
Joined: 15 May 2008
Posts: 265
Location: Lawrence, KS
Posted: Mon Mar 14, 2011 8:48 am Post subject:
Daniel: great work, as always. Are you planning to update your numbers through Sunday's games? (Looks like they're current through last Friday, so I suspect the update would only make a minor difference for most teams.) It will be interesting to see how your model fares in predicting upcoming NCAA tourney games.
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 10:11 am Post subject:
Ilardi wrote:
Daniel: great work, as always. Are you planning to update your numbers through Sunday's games? (Looks like they're current through last Friday, so I suspect the update would only make a minor difference for most teams.) It will be interesting to see how your model fares in predicting upcoming NCAA tourney games.
I will, yes. Unless, of course, my site is crashed by excessive traffic. It seems to be down now..._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 10:23 am Post subject:
The latest Bayesian Ratings are here, in an Excel sheet: http://bit.ly/dI2fza_________________Go ... com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Mon Mar 14, 2011 11:57 am Post subject:
Daniel, thanks for explaining that the Bayesian column represents 100 possessions. I'm referring to a scale that would show all the teams at once in terms of how they relate in point differential, as Jeff Sagarin's been doing at USA Today for eons for example. http://www.usatoday.com/sports/sagarin/bkt1011.htm Would you recommend multiplying the Bayesian column by .667 (eyeballing the midpoint pace factor at kenpom) to approximate 66.7 possessions per game, then using that as a standard for a 68-team comparison in the tourney? I'm aware that it's not as ideal as running each and every conceivable matchup through the algebra ringer...but that's a bit much if you're just trying to see how the teams in a certain regional stack up against each other, etc... First used the formula you sited for basketball projections back in 1984 with simple game total averages in the NBA. Got the job done well for Over/Unders back then. Oddsmakers didn't realize how much things would blow up or blow down when extremes played each other. Formula captured it very well. When expanded boxscores became widely available, we used half of free throw attempts rather than .44 in the possession estimates. Ahead of the curve I guess, but not quite as exact as it could have been. Nowadays in the colleges many offshore places reportedly just use kenpom's "Fanmatch" page for posting their over/unders. Saves them a lot of work. Not the same with his game margins though. Less agreement about teams there. Remember kenpom talking in a blog several weeks back about readers being less than enthusiastic about his game/margin predictions. Not able to outperform the prediction markets yet. Maybe soon. Seems very close but the shadings aren't quite there yet I guess.
Back to top
EvanZ
Joined: 22 Nov 2010
Posts: 290
Posted: Mon Mar 14, 2011 1:25 pm Post subject:
Seems to me like using Daniel's rankings straight up (or something like LRMC) is the way to go to improve chances of getting 2nd or 3rd place in a pool, but may not be the best way to actually win a pool. Above and beyond the "upsets" that the (presumably more accurate) Bayesian rankings predict, there will be upsets that could not possibly be predicted, except by random chance. I don't see someone winning a relatively large sized pool without making some "crazy" picks, if that makes sense._________________http://www.thecity2.com http://www.ibb.gatech.edu/evan-zamir
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 2:25 pm Post subject:
EvanZ wrote:
Seems to me like using Daniel's rankings straight up (or something like LRMC) is the way to go to improve chances of getting 2nd or 3rd place in a pool, but may not be the best way to actually win a pool. Above and beyond the "upsets" that the (presumably more accurate) Bayesian rankings predict, there will be upsets that could not possibly be predicted, except by random chance. I don't see someone winning a relatively large sized pool without making some "crazy" picks, if that makes sense.
Right. The way to maximize your rank, though, in large pools, is not to go with the Bayesian ratings if you know the distribution of selections. Since we know what was picked in the ESPN pool, we can choose using that information. Maximize RoundValue*(Odds% + (Odds%- Chosen%)). In other words, pick against the crowd, but don't pick teams that don't have any chance._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
EvanZ
Joined: 22 Nov 2010
Posts: 290
Posted: Mon Mar 14, 2011 2:27 pm Post subject:
DSMok1 wrote:
EvanZ wrote:
Seems to me like using Daniel's rankings straight up (or something like LRMC) is the way to go to improve chances of getting 2nd or 3rd place in a pool, but may not be the best way to actually win a pool. Above and beyond the "upsets" that the (presumably more accurate) Bayesian rankings predict, there will be upsets that could not possibly be predicted, except by random chance. I don't see someone winning a relatively large sized pool without making some "crazy" picks, if that makes sense.
Right. The way to maximize your rank, though, in large pools, is not to go with the Bayesian ratings if you know the distribution of selections. Since we know what was picked in the ESPN pool, we can choose using that information. Maximize RoundValue*(Odds% + (Odds%- Chosen%)). In other words, pick against the crowd, but don't pick teams that don't have any chance.
Rhetorical question...can you come up with a simulation that would tell you how risky to be depending on pool size? That would be pretty nifty._________________http://www.thecity2.com http://www.ibb.gatech.edu/evan-zamir
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Mon Mar 14, 2011 4:39 pm Post subject:
For people thinking about office pools... Kenpom (current offshore line in parenthesis) Thursday http://kenpom.com/fanmatch.php?d=2011-03-17 Wisconsin by 3 (4.5) Cincinnati by 2 (1) Gonzaga by 1 (St. John's by 1.5) Utah State by 3 (K-State by 2) Vandy by 2 (2) Temple by 1 (2.5) Michigan State by 1 (1.5) ODU by 1 (2) BYU by 13 (8.5) UCONN by 9 (10) Kentucky by 12 (13.5) Louisville by 12 (10) Florida by 12 (12.5) San Diego State by 15 (15.5) Friday http://kenpom.com/fanmatch.php?d=2011-03-18 Illinois by 1 (UNLV by 1.5) G. Mason by 1 (1.5) Marquette by 1 (Xavier by 2) FSU by 1 (pick) Michigan by 2 (Tennessee by 1.5) Washington by 6 (5.5) Arizona by 8 (6) Texas by 14 (10) Syracuse by 13 (11.5) N. Carolina by 18 (17.5) Notre Dame by 17 (13) Purdue by 15 (14) Kansas by 22 (22.5) Duke by 27 (22.5) Many virtual coin flips in the first round, with some disagreement between kenpom and the market in those. Not many methodologies get the coin flip games right when you have to pick them ALL in a pool (lol). As you guys point out, you can create value potential by going opposite the masses if the masses line up on one side of a coin flip... In some seasons, I think spreads up to as high as 3-4 ish have ended up splitting out as if they were true coin flips. That will happen with this kind of sample size though. Tough to know teams with "certainty" even at this point of the season given strength of schedule issues, injury/suspension issues, young teams getting better as they mature, etc... Market less enthusiastic about Texas than kenpom is...
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Tue Mar 15, 2011 2:50 pm Post subject:
Thought there might be some interest in this work by some Stanford guys, as referenced in Chad Millman's ESPN.com article on the tourney. An explanation of their models: http://www.teamrankings.com/blog/ncaa-b ... everywhere These look to be their team rankings heading in. Not familiar with them at all, so please don't think of this as vouching for their data or their marketing. Just saw the references in Millman's article, and thought the explanation of their modeling process might interest people who are thinking about modeling... http://www.teamrankings.com/ncaa-basket ... ng-by-team
Author Message DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 5:17 pm Post subject: On Bayesian Predictive Efficiency Rankings
I just put up a tremendously long post on my website discussing how to use Bayesian updating with NBA predictive efficiency ratings. I'm sure there are some flaws (such as how to adjust for opponent), so let me know! http://godismyjudgeok.com/DStats/2011/n ... com/DStats Twitter.com/DSMok1
Back to top
back2newbelf
Joined: 21 Jun 2005
Posts: 274
Posted: Tue Feb 15, 2011 5:27 pm Post subject:
I get an Error in query: User not signed in <a target="_blank" href="https://spreadsheets0.google.com/">Sign in</a> and the last charts don't show up_________________http://stats-for-the-nba.appspot.com/
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 5:36 pm Post subject:
back2newbelf wrote:
I get an Error in query: User not signed in <a target="_blank" href="https://spreadsheets0.google.com/">Sign in</a> and the last charts don't show up
Try again--does it work now?_________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
back2newbelf
Joined: 21 Jun 2005
Posts: 274
Posted: Tue Feb 15, 2011 5:50 pm Post subject:
Yes it works now_________________http://stats-for-the-nba.appspot.com/
Back to top
Ilardi
Joined: 15 May 2008
Posts: 265
Location: Lawrence, KS
Posted: Tue Feb 15, 2011 6:38 pm Post subject:
DSM - Intriguing work, as usual. So, I'm curious to know if your Bayesian model does a reasonable job approximating the observed distribution of actual game results? In other words, if we took a largish historical sample of n games, and for each game computed your model's predicted game outcome (based on the Bayesian efficiency differential of each team's antecedent games), would the actual observed game outcomes be distributed more or less normally, with the predicted distribution means and sd's? Steve
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 6:57 pm Post subject:
Ilardi wrote:
DSM - Intriguing work, as usual. So, I'm curious to know if your Bayesian model does a reasonable job approximating the observed distribution of actual game results? In other words, if we took a largish historical sample of n games, and for each game computed your model's predicted game outcome (based on the Bayesian efficiency differential of each team's antecedent games), would the actual observed game outcomes be distributed more or less normally, with the predicted distribution means and sd's? Steve
Sounds like another post! _________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Crow
Joined: 20 Jan 2009
Posts: 821
Posted: Tue Feb 15, 2011 9:07 pm Post subject:
Do you want to predict the playoffs with this tool and compare to Hollinger's and the others out there?
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Tue Feb 15, 2011 11:33 pm Post subject:
Crow wrote:
Do you want to predict the playoffs with this tool and compare to Hollinger's and the others out there?
Not really. I'd prefer to predict the playoffs based on Bayesian-based ASPM._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 70
Posted: Wed Feb 16, 2011 12:44 am Post subject:
Not sure if you would consider this helpful or not DSM, but I know a programmer who used Bayesian priors for all boxscore stats in a predictive format two seasons ago (and had a methodology for ranking the importance of each stat based on its relationship to winning as the season played itself out). If I recall, he was surprised to find that the most predictive hunk of games was the prior five...rather than much longer term samples. He had expected larger sample sizes to make better predictions. Ultimately he decided the approach was still inferior to the prediction markets reflected by the Vegas pointspreads. Those have the ability to react to on the fly information like injuries, magnified fatigue spots, guys coming back from injuries, etc... Best of luck with your efforts...
Back to top
bbstats
Joined: 25 Apr 2010
Posts: 46
Posted: Wed Feb 16, 2011 12:48 am Post subject:
You had me at "bayesian." Great stuff._________________http://thebasketballdistribution.blogspot.com http://twitter.com/bbstats
Back to top
bbstats
Joined: 25 Apr 2010
Posts: 46
Posted: Wed Feb 16, 2011 1:08 am Post subject:
PS - Would I be correct in saying that the theory here is similar to Dean Oliver's Kalman Filter?_________________http://thebasketballdistribution.blogspot.com http://twitter.com/bbstats
Back to top
EvanZ
Joined: 22 Nov 2010
Posts: 295
Posted: Wed Feb 16, 2011 6:34 am Post subject:
bbstats wrote:
You had me at "bayesian." Great stuff.
He had me at "On...". Any time you see that leading preposition, you know some mathematical wizardry is to follow. BTW, has anyone used bootstrapping to do something similar? Or maybe that could be useful to create a prior?_________________http://www.thecity2.com http://www.ibb.gatech.edu/evan-zamir
Back to top
John Hollinger
Joined: 14 Feb 2005
Posts: 175
Posted: Wed Feb 16, 2011 10:48 am Post subject:
Fascinating stuff, and interesting that it supports my notion that Philly and Memphis are both a hell of a lot better than people realize and Utah is much worse. I suspect our big difference on Dallas is due to the timing of Dirks injury since I just weight last 10 rather than using a gradual function like you did....
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Wed Feb 16, 2011 11:23 am Post subject:
John Hollinger wrote:
Fascinating stuff, and interesting that it supports my notion that Philly and Memphis are both a hell of a lot better than people realize and Utah is much worse. I suspect our big difference on Dallas is due to the timing of Dirks injury since I just weight last 10 rather than using a gradual function like you did....
When you do the time-weighting, how & when in the calculations do you adjust for opponents? In this analysis, I first ran fully adjusted team efficiency differentials (over the whole season, w/o time-weighting), and used that value to pre-adjust the game efficiency differentials, before doing the Bayesian time-weighting. How do you adjust for opponents in the "last 10" component? I'm puzzling how to do that part of the analysis more rigorously._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 70
Posted: Wed Feb 16, 2011 1:12 pm Post subject:
Regarding JH's notes on Philly/Memphis/Utah, confirmation from the markets... *Philadelphia has topped expectations to the tune of 34-20-1 this season, including 29-14 the last 43 games. *Memphis is 35-21-1, including 28-12 the last 40 games. *Utah is 25-30-1 for the year, including 8-19 the last 27 games. If one accepts the premise that market prices are a composite of "what people are thinking," we can probably move that from notion to confirmed fact I'd think.
Author Message Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Thu Feb 17, 2011 12:26 am Post subject:
Updating to 35-20-1 for Philly vs. expectations, 30-14 the last 44 heading into the ASB after Wednesday's win in Houston... And, 25-31-1 for Utah, 8-20 the last 28 heading into the ASB after Wednesday's home loss to Golden State.
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Fri Feb 18, 2011 5:47 pm Post subject:
Now, with NCAA ratings! I also put up a spreadsheet which you all may find very interesting to manipulate._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 07, 2011 7:44 pm Post subject:
DSMok1 wrote:
Now, with NCAA ratings! I also put up a spreadsheet which you all may find very interesting to manipulate.
I've updated this data and added some more things. I still haven't gotten around to doing the standard errors for the NCAA and verifying the errors with game results, sorry. http://godismyjudgeok.com/DStats/2011/n ... com/DStats Twitter.com/DSMok1
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Sat Mar 12, 2011 10:10 pm Post subject:
And another new version, this time incorporating previous year information as a better-informed Bayesian prior: http://godismyjudgeok.com/DStats/2011/n ... com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Sun Mar 13, 2011 11:43 am Post subject:
DSM, would you consider the numbers in your Bayesian column to be a reasonable estimation of the point differences between teams in a 40 minute game? Meaning, Kansas is about 3 points better on a neutral court than Texas because they're 3 points higher? If not, is there a way to easily convert your output into something that would resemble a point differential scale so readers could compare it to the market prices that go up in the NCAA Tournament? Living in Austin, it was cool to see Texas so high. Wish they could find a high level of form more consistently. I'm afraid they peaked too early again...
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 6:29 am Post subject:
Jeff Fogle wrote:
DSM, would you consider the numbers in your Bayesian column to be a reasonable estimation of the point differences between teams in a 40 minute game? Meaning, Kansas is about 3 points better on a neutral court than Texas because they're 3 points higher? If not, is there a way to easily convert your output into something that would resemble a point differential scale so readers could compare it to the market prices that go up in the NCAA Tournament? Living in Austin, it was cool to see Texas so high. Wish they could find a high level of form more consistently. I'm afraid they peaked too early again...
Those are points per 100 possessions, not per game. To calculate the number of possessions expected in a game, take adjusted pace from Pomeroy for each team, and calculate as PaceA*PaceB/NCAAAvg. to get the expected pace for each game._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
Ilardi
Joined: 15 May 2008
Posts: 265
Location: Lawrence, KS
Posted: Mon Mar 14, 2011 8:48 am Post subject:
Daniel: great work, as always. Are you planning to update your numbers through Sunday's games? (Looks like they're current through last Friday, so I suspect the update would only make a minor difference for most teams.) It will be interesting to see how your model fares in predicting upcoming NCAA tourney games.
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 10:11 am Post subject:
Ilardi wrote:
Daniel: great work, as always. Are you planning to update your numbers through Sunday's games? (Looks like they're current through last Friday, so I suspect the update would only make a minor difference for most teams.) It will be interesting to see how your model fares in predicting upcoming NCAA tourney games.
I will, yes. Unless, of course, my site is crashed by excessive traffic. It seems to be down now..._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 10:23 am Post subject:
The latest Bayesian Ratings are here, in an Excel sheet: http://bit.ly/dI2fza_________________Go ... com/DStats Twitter.com/DSMok1
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Mon Mar 14, 2011 11:57 am Post subject:
Daniel, thanks for explaining that the Bayesian column represents 100 possessions. I'm referring to a scale that would show all the teams at once in terms of how they relate in point differential, as Jeff Sagarin's been doing at USA Today for eons for example. http://www.usatoday.com/sports/sagarin/bkt1011.htm Would you recommend multiplying the Bayesian column by .667 (eyeballing the midpoint pace factor at kenpom) to approximate 66.7 possessions per game, then using that as a standard for a 68-team comparison in the tourney? I'm aware that it's not as ideal as running each and every conceivable matchup through the algebra ringer...but that's a bit much if you're just trying to see how the teams in a certain regional stack up against each other, etc... First used the formula you sited for basketball projections back in 1984 with simple game total averages in the NBA. Got the job done well for Over/Unders back then. Oddsmakers didn't realize how much things would blow up or blow down when extremes played each other. Formula captured it very well. When expanded boxscores became widely available, we used half of free throw attempts rather than .44 in the possession estimates. Ahead of the curve I guess, but not quite as exact as it could have been. Nowadays in the colleges many offshore places reportedly just use kenpom's "Fanmatch" page for posting their over/unders. Saves them a lot of work. Not the same with his game margins though. Less agreement about teams there. Remember kenpom talking in a blog several weeks back about readers being less than enthusiastic about his game/margin predictions. Not able to outperform the prediction markets yet. Maybe soon. Seems very close but the shadings aren't quite there yet I guess.
Back to top
EvanZ
Joined: 22 Nov 2010
Posts: 290
Posted: Mon Mar 14, 2011 1:25 pm Post subject:
Seems to me like using Daniel's rankings straight up (or something like LRMC) is the way to go to improve chances of getting 2nd or 3rd place in a pool, but may not be the best way to actually win a pool. Above and beyond the "upsets" that the (presumably more accurate) Bayesian rankings predict, there will be upsets that could not possibly be predicted, except by random chance. I don't see someone winning a relatively large sized pool without making some "crazy" picks, if that makes sense._________________http://www.thecity2.com http://www.ibb.gatech.edu/evan-zamir
Back to top
DSMok1
Joined: 05 Aug 2009
Posts: 611
Location: Where the wind comes sweeping down the plains
Posted: Mon Mar 14, 2011 2:25 pm Post subject:
EvanZ wrote:
Seems to me like using Daniel's rankings straight up (or something like LRMC) is the way to go to improve chances of getting 2nd or 3rd place in a pool, but may not be the best way to actually win a pool. Above and beyond the "upsets" that the (presumably more accurate) Bayesian rankings predict, there will be upsets that could not possibly be predicted, except by random chance. I don't see someone winning a relatively large sized pool without making some "crazy" picks, if that makes sense.
Right. The way to maximize your rank, though, in large pools, is not to go with the Bayesian ratings if you know the distribution of selections. Since we know what was picked in the ESPN pool, we can choose using that information. Maximize RoundValue*(Odds% + (Odds%- Chosen%)). In other words, pick against the crowd, but don't pick teams that don't have any chance._________________GodismyJudgeOK.com/DStats Twitter.com/DSMok1
Back to top
EvanZ
Joined: 22 Nov 2010
Posts: 290
Posted: Mon Mar 14, 2011 2:27 pm Post subject:
DSMok1 wrote:
EvanZ wrote:
Seems to me like using Daniel's rankings straight up (or something like LRMC) is the way to go to improve chances of getting 2nd or 3rd place in a pool, but may not be the best way to actually win a pool. Above and beyond the "upsets" that the (presumably more accurate) Bayesian rankings predict, there will be upsets that could not possibly be predicted, except by random chance. I don't see someone winning a relatively large sized pool without making some "crazy" picks, if that makes sense.
Right. The way to maximize your rank, though, in large pools, is not to go with the Bayesian ratings if you know the distribution of selections. Since we know what was picked in the ESPN pool, we can choose using that information. Maximize RoundValue*(Odds% + (Odds%- Chosen%)). In other words, pick against the crowd, but don't pick teams that don't have any chance.
Rhetorical question...can you come up with a simulation that would tell you how risky to be depending on pool size? That would be pretty nifty._________________http://www.thecity2.com http://www.ibb.gatech.edu/evan-zamir
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Mon Mar 14, 2011 4:39 pm Post subject:
For people thinking about office pools... Kenpom (current offshore line in parenthesis) Thursday http://kenpom.com/fanmatch.php?d=2011-03-17 Wisconsin by 3 (4.5) Cincinnati by 2 (1) Gonzaga by 1 (St. John's by 1.5) Utah State by 3 (K-State by 2) Vandy by 2 (2) Temple by 1 (2.5) Michigan State by 1 (1.5) ODU by 1 (2) BYU by 13 (8.5) UCONN by 9 (10) Kentucky by 12 (13.5) Louisville by 12 (10) Florida by 12 (12.5) San Diego State by 15 (15.5) Friday http://kenpom.com/fanmatch.php?d=2011-03-18 Illinois by 1 (UNLV by 1.5) G. Mason by 1 (1.5) Marquette by 1 (Xavier by 2) FSU by 1 (pick) Michigan by 2 (Tennessee by 1.5) Washington by 6 (5.5) Arizona by 8 (6) Texas by 14 (10) Syracuse by 13 (11.5) N. Carolina by 18 (17.5) Notre Dame by 17 (13) Purdue by 15 (14) Kansas by 22 (22.5) Duke by 27 (22.5) Many virtual coin flips in the first round, with some disagreement between kenpom and the market in those. Not many methodologies get the coin flip games right when you have to pick them ALL in a pool (lol). As you guys point out, you can create value potential by going opposite the masses if the masses line up on one side of a coin flip... In some seasons, I think spreads up to as high as 3-4 ish have ended up splitting out as if they were true coin flips. That will happen with this kind of sample size though. Tough to know teams with "certainty" even at this point of the season given strength of schedule issues, injury/suspension issues, young teams getting better as they mature, etc... Market less enthusiastic about Texas than kenpom is...
Back to top
Jeff Fogle
Joined: 11 Jan 2011
Posts: 68
Posted: Tue Mar 15, 2011 2:50 pm Post subject:
Thought there might be some interest in this work by some Stanford guys, as referenced in Chad Millman's ESPN.com article on the tourney. An explanation of their models: http://www.teamrankings.com/blog/ncaa-b ... everywhere These look to be their team rankings heading in. Not familiar with them at all, so please don't think of this as vouching for their data or their marketing. Just saw the references in Millman's article, and thought the explanation of their modeling process might interest people who are thinking about modeling... http://www.teamrankings.com/ncaa-basket ... ng-by-team